jyasskin
commented
1 year ago (I renamed the issue to avoid code-of-conduct problems, which I don't think Martin really intended)
Websites often need to know which legal jurisdiction a user is in, which is low-granularity geolocation. Am I still naive to think that that level of granularity is "less sensitive for many people"? We might be missing a general warning that it's harder than expected to be sure that information is actually gone when you try to obscure it, which would apply to de-identification too. For legal jurisdiction, for example, a transition between jurisdictions reveals that the person was precisely on the border at some point during the transition.
npdoty
As with many things privacy-related, the naive notion is often badly wrong.
See Section 2.4:
Section 13.5 of RFC 6772 is something I refer people to often in these cases. This highlights how an intuition about "precision" can be badly wrong and how apparent fuzziness in data can hide a great deal of information. (This is also true for de-identification, which is routinely done poorly; but I'll open other issues about that.)
A good example is a paper written at approximately the same time that was able to recover geolocation at near-GPS levels of precision, using only location data from cell towers combined with knowledge of the terrain (I tried to dig up a reference, but was not successful; it's been a while; I wish that we'd manage to cite it). I have no idea what modern machine learning techniques would make of that problem, by the way.